On the convexity of piecewise-defined functions

Transcription

1 On the convexity of piecewise-defined functions arxiv: v1 [math.ca] 16 Aug 2014 Heinz H. Bauschke, Yves Lucet, and Hung M. Phan August 16, 2014 Abstract Functions that are piecewise defined are a common sight in mathematics while convexity is a property especially desired in optimization. Suppose now a piecewisedefined function is convex on each of its defining components when can we conclude that the entire function is convex? In this paper we provide several convenient, verifiable conditions guaranteeing convexity (or the lack thereof). Several examples are presented to illustrate our results Mathematics Subject Classification: Primary 26B25; Secondary 52A41, 65D17, 90C25. Keywords: computer-aided convex analysis, convex function, convex interpolation, convex set, piecewise-defined function. 1 Introduction Consider the function x 2 + y max{0, xy} (1) f : R 2, if(x, y) = (0, 0); R : (x, y) x + y 0, otherwise. Clearly, f is a piecewise-defined function with continuous components (2a) f 1 (x, y) := x+y on A 1 := R + R + ; Mathematics, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. heinz.bauschke@ubc.ca. Computer Science, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. yves.lucet@ubc.ca Mathematics, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. hung.phan@ubc.ca. 1

2 (2b) (2c) (2d) f 2 (x, y) := x2 + y 2 x+y on A 2 := R R + ; f 3 (x, y) := x+y on A 3 := R R ; f 4 (x, y) := x2 + y 2 x y on A 4 := R + R. One may check that each f i is a convex function (see Example 6.1 below for details). However, whether or not f itself is convex is not immediately clear. (As it turns out, f is convex.) On the other hand, if (3a) (3b) f 1 (x) = x on A 1 := R ; and f 2 (x) = x on A 2 := R +, then f 1 and f 2 are convex while the induced piecewise-defined function f(x) = x is not convex. These and similar examples motivate the goal of this paper which is to present verifiable conditions guaranteeing the convexity of a piecewise-defined function provided that each component is convex. Special cases of our results have been known in the convex interpolation community (see Remark 5.6). Moreover, our results have applications to computer-aided convex analysis (see Remark 5.8). The remainder of this paper is organized as follows. In Section 2, we collect various auxiliary results concerning convexity and differentiability. We also require properties of collections of sets and of functions which we develop in Section 3 and Section 4, respectively. Our main results guaranteeing convexity are presented in Section 5. Various examples illustrating convexity and the lack thereof are discussed in Section 6 and Section 7, respectively. Notation: Throughout, X is a Euclidean space with inner product, and induced norm. R denotes the set of real numbers, R + := { x R } x 0, and R := R +. For x and y in X, [x, y] := { (1 t)x+ty } 0 t 1 is the line segment connecting x and y. Similarly, we set]x, y[ := { (1 t)x+ty } { } 0 < t < 1,[x, y[ := (1 t)x+ ty 0 t < 1, and ]x, y] := [y, x[. For a subset A of X, conv A, cl A, int A, aff A, and ri A respectively denote the convex hull, the closure, the interior, the affine hull, and the relative interior of A. Furthermore, ι A is the indicator function of A defined by ι A (x) = 0, if x A; and + otherwise. Let f : X ],+ ] = R {+ }. The domain of f is D f := { x X } f(x) < + ; f is said to be proper if D f =. The restriction of f on some subset A of X is denoted by f A. A set-valued mapping F from X to another Euclidean space Y is denoted by F : X Y; and its domain is D F := { x X } F(x) =. For further background and notation, we refer the reader to [1, 9, 10, 11, 14]. 2

3 2 Convexity and differentiability Let f : X ],+ ] be proper. For every x X, the subdifferential (in the sense of convex analysis) of f at x, denoted by f(x), is the set of all vectors x X such that (4) ( y X) x, y x f(y) f(x). The induced operator f : X X has domain D f = { x X f(x) = } D f. Let us now present some auxiliary results concerning the convexity of a function. Lemma 2.1 Let f : X ],+ ] and let z 1 and z 2 be in D f. Set x := (1 t)z 1 + tz 2, where t [0, 1], and assume that x D f. Then f(x) (1 t) f(z 1 )+t f(z 2 ). Proof. Let x f(x). Then x, z 1 x f(z 1 ) f(x) and x, z 2 x f(z 2 ) f(x). Hence (5a) (5b) (1 t) x, z 1 x (1 t)( f(z 1 ) f(x)); t x, z 2 x t( f(z 2 ) f(x)). Adding up the last two inequalities, we obtain 0 (1 t) f(z 1 )+t f(z 2 ) f(x). Fact 2.2 (See [10, Theorem ].) Let A be a nonempty convex subset of X. following hold: Then the (i) ( x cl A)( y ri A) ]x, y] ri A. (ii) ri A is nonempty and convex. (iii) cl(ri A) = cl A. Fact 2.3 (See [10, Theorem 23.4].) Let f : X ],+ ] be convex and proper. Then ri D f D f. Fact 2.4 (See [14, Theorem 2.4.1(iii)]) Let f : X ],+ ] be proper. Assume that D f = D f is convex. Then f is convex. Proof. Take z 1 and z 2 in D f and let t [0, 1]. Set x := (1 t)z 1 + tz 2. Since f(x) =, Lemma 2.1 implies that f(x) (1 t) f(z 1 )+t f(z 2 ). Therefore, f is convex. In the presence of continuity, Fact 2.4 admits the following extension. Lemma 2.5 Let f : X ],+ ] be proper. Assume that D f is convex, that f D f is continuous, and that ri D f D f. Then f is convex. 3

4 Proof. Fact 2.2 implies that ri D f is convex and cl(ri D f ) = cl D f. Then the function f + ι ri D f is convex by Fact 2.4. Since f D f is continuous and ri D f is dense in D f, we conclude that f D f is convex. Given a nonempty subset A of X, we define the dimension of A to be the dimension of the linear subspace parallel to the affine hull of A, i.e., dim A := dim(aff A aff A). We then have the following result. Lemma 2.6 Let f : X ],+ ] be proper. Assume that f D is continuous, that D f is f convex and at least 2-dimensional, and that there exists a finite subset E of X such that f [x,y] is convex for every segment[x, y] contained in(ri D f ) E. Then f is convex. Proof. Take two distinct points x and y in D f, let t [0, 1], and set z := (1 t)x+ty. Then z D f because D f is convex. It remains to show that (6) f(z) (1 t) f(x)+ t f(y). First, since D f is convex and dim(d f ) 2, there exists w (ri D f ) aff{x, y}. For each ε ]0, 1[, set (7) x ε := (1 ε)x+εw, y ε := (1 ε)y+ εw, and z ε := (1 t)x ε + ty ε. Using Fact 2.2(i) and the finiteness of E, we have (8) z ε [x ε, y ε ] (ri D f ) E, for every ε > 0 sufficiently small. Because f is convex on [x ε, y ε ], this implies (9) f(z ε ) (1 t) f(x ε )+t f(y ε ). Letting ε 0 +, we obtain (6) by using the continuity of f D f. Remark 2.7 (the assumption on the dimension is important) Lemma 2.6 fails on R in the following sense. Consider f : R R : x x and set E := {0}. Then all assumptions of Lemma 2.6 hold except that D f = R is only 1-dimensional. Clearly, the conclusion of Lemma 2.6 is not true because f is not convex. We now turn our attention to differentiability properties. Recall that f : X ],+ ] is differentiable at x int D f if there exists f(x) X such that ( y X) f(y) f(x) f(x), y x = o( y x ); f is differentiable on subset A of int D f if f is differentiable at every x A. We will require the following results. Fact 2.8 (See [10, Theorem 25.1].) Let f : X ],+ ] be convex and proper, and assume that x int D f. Then f is differentiable at x if and only if f(x) is a singleton. 4

5 Fact 2.9 (See [10, Theorem 25.5].) Let f : X ],+ ] be convex and proper, and let Ω be the set of points where f is differentiable. Then Ω is a dense subset of int D f, and its complement in int D f is a set of measure zero. Moreover, f : Ω X is continuous. Fact 2.10 (See [10, Theorem 25.6].) Let f : X ],+ ] be convex and proper such that D f is closed with nonempty interior. Then (10) ( x D f ) f(x) = cl(conv S(x))+N D f (x), where N D f (x) := { x X ( y D f ) x, y x 0 } is the normal cone to D f at x and S(x) is the set of all limits of sequences( f(x n )) n N such that f is differentiable at every x n and x n x. 3 Compatible systems of sets In this section, we always assume that (11a) (11b) (11c) I is a nonempty finite set; A := {A i } i I is a system of convex subsets of X; A := i I A i. Definition 3.1 (compatible systems of sets) Assume (11). We say that A is a compatible system of sets if i I (12) j I cl A i cl A j ri A = A i A j ri A; i = j otherwise, we say thatais incompatible. Example 3.2 Every system of finitely many closed convex subsets of X is compatible. Example 3.3 (incompatible systems) Suppose that X = R 2, that I = {1, 2}, that A 1 = ]0, 1] [0, 1], and that A 2 = [ 1, 0] [0, 1]. Then A = A 1 A 2 = [ 1, 1] [0, 1] and ri A = ] 1, 1[ ]0, 1[. Thus, A = {A 1, A 2 } is incompatible because (13) cl A 1 cl A 2 ri A = {0} ]0, 1[ = = A 1 A 2 ri A. Definition 3.4 (colinearly ordered tuple) The tuple of vectors(x 0,..., x n ) X n is said to be colinearly ordered if the following hold: (i) [x 0, x n ] = [x 0, x 1 ] [x n 1, x n ]; (ii) 0 x 0 x 1 x 0 x 2 x 0 x n. 5

6 Proposition 3.5 Assume (11) and that A is a compatible system of sets (recall Definition 3.1). Then for every segment[x, y] contained in ri A, there exists a colinearly ordered tuple(x 0,..., x n ) and {A i1,..., A in } A such that ( ) (14) x 0 = x; x n = y; and k {1,..., n} [x k 1, x k ] A ik. Proof. Let [x, y] ri A, with x A i1 for some i 1 I. Set x 0 := x. For every t [0, 1], define (15) x(t) := (1 t)x+ty. Furthermore, set (16) t 1 := sup { t [0, 1] x(t) Ai1 } and x1 := x(t 1 ). Then [x 0, x 1 [ A i1 and x 1 cl A i1. Note also that x 1 ri A. Case 1: t 1 = 1. Then x 1 = y cl A i1. Suppose that y A i1. Then, y A i2 for some i 2 I. It follows that y cl A i1 cl A i2 ri A = A i1 A i2 ri A, which is a contradiction. Therefore, y A i1 and we are done because [x, y] A i1. Case 2: t 1 < 1. Then there exist ε ]0, 1 t 1 ] and A i2 A {A i1 } such that (17) ]x 1, x 2 ] A i2 where x 2 := x(t 1 + ε). Hence x 1 cl A i2. We then have (18) x 1 cl A i1 cl A i2 ri A = A i1 A i2 ri A. So we have split [x, y] into two line segments (19) [x 0, x 1 ] A i1 ri A and [x 1, y] ( i I {i 1 } A i ) ri A. Next, we repeat the above process for the segment[x 1, y]. SinceAis finite, we eventually obtain (14). Remark 3.6 (closedness is not necessary for compatibility) We note that there are compatible systems of sets that are not closed. For example, suppose that X = R 2, that I = {1, 2}, that A 1 = [0, 1[ [0, 1], and that A 2 = ] 1, 0] [0, 1]. Then neither A 1 nor A 2 is closed. However, since (20) cl A 1 cl A 2 = A 1 A 2 = {0} [0, 1], we deduce that A = {A 1, A 2 } is compatible. Definition 3.7 (active index set) Assume (11). For every x X, we define the active index set associated with A by (21) I A (x) := { i I x Ai }, and we will write I(x) if there is no cause for confusion. 6

7 Proposition 3.8 Assume (11) and that A is a compatible system of sets (see Definition 3.1). Suppose that x int A and that I A (x) = {i}. Then x int A i. Proof. Because int A =, we have ri A = int A. Suppose to the contrary that x int A i. Then there exist j I {i} and a sequence(x n ) n N in A j such that that x n x. It follows that (22) x cl A i cl A j ri A = A i A j ri A, which is absurd because I(x) = {i} by assumption. 4 Compatible systems of functions In this section, we always assume that (23a) (23b) (23c) (23d) I is a nonempty finite set; F := { f i } i I is a system of proper convex functions from X to ],+ ]; f := min i I f i is the piecewise-defined function associated with F; I F : X I : x { i I } x D fi is the active index set function. We will write I(x) instead of I F (x) if there is no cause for confusion. Note that D f = i I D fi. Definition 4.1 (compatible systems of functions) Assume (23). We say that F is a compatible system of functions if( i I) f i D fi is continuous and (24) i I j I f i D f D fi D j. fj fi D fj i = j We start with a useful lemma. Lemma 4.2 Assume (23) and that F is compatible system of functions (recall Definition 4.1). Then (25) ( x X) f(x) i I F (x) f i (x). Proof. Suppose that x f(x) and that i I F (x). Then f i (x) = f(x) and ( y X) f i (y) f i (x) f(y) f(x) x, y x. Therefore, x f i (x). 7

8 Lemma 4.3 Let (a, b, c) X 3 be colinearly ordered (recall Definition 3.4). Assume (23) with I = {1, 2}, thatf is compatible system of functions (recall Definition 4.1), that D f1 = [a, b], that D f2 = [b, c], and that (26) f 1 (b) f 2 (b) =. Then f is convex and (27) ( x D f ) f(x) = i I F (x) f 1 (x), if x [a, b[ ; f i (x) = f 1 (b) f 2 (b), if x = b; f 2 (x), if x ]b, c]. Proof. We assume that a, b, c are pairwise distinct since the other cases are trivial. First, we show that ( ) (28) x [a, b[ f 1 (x) f(x), Suppose that x [a, b[ and that x f 1 (x). To establish (28), it suffices to show that ( ) (29) y [a, c] x, y x f(y) f(x). Indeed, (29) is true for y [a, b] by definition of f 1 (x) and f. Now suppose that y ]b, c]. By (26), there exists b f 1 (b) f 2 (b). Then (30a) (30b) (30c) (30d) (30e) (30f) (30g) x, y x = x, b x + x, y b f 1 (b) f 1 (x)+ y b b x x, b x f 1 (b) f 1 (x)+ y b b x b, b x f 1 (b) f 1 (x)+ b, y b f 1 (b) f 1 (x)+ f 2 (y) f 2 (b) = f 2 (y) f 1 (x) = f(y) f(x). Hence (29) holds, as does (28). Switching the roles of f 1 and f 2, we obtain analogously (31) ( x [c, b[) f 2 (x) f(x). Next, it is straightforward to check that (32) f 1 (b) f 2 (b) f(b). Since the reverse inclusions of (28), (31), and (32) follow from Lemma 4.2, we conclude that (27) holds. Using (26), (27) and Fact 2.3, we conclude that f(x) = for all x ]a, c[ = ri D f. Finally, it follows from Lemma 2.5 that f is convex. 8

9 Theorem 4.4 Let(x 0,..., x n ) X n+1 be colinearly ordered (recall Definition 3.4). Assume (23) with I = {1,..., n} and that F is a compatible system of functions (recall Definition 4.1) such that the following hold: (i) ( i {1,..., n}) D fi = [x i 1, x i ]. (ii) ( i {1,..., n 1}) f i (x i ) f i+1 (x i ) =. Then f is convex and (33) ( x D f ) f(x) = f i (x); i I F (x) Proof. If n = 1, then the result is trivial. For n 2, the result follows by inductively applying Lemma Main results We are now ready for our main results. Theorem 5.1 (main result I) Assume (23), that F is a compatible system of functions (recall Definition 4.1), and that the following hold: (a) D f = i I D fi is convex and at least 2-dimensional. (b) {D fi } i I is a compatible system of sets (recall Definition 3.1). (c) There exists a finite subset E of X such that } x (ri D (34) f ) E card I(x) 2 i I(x) f i (x) =. Then f is convex and (35) ( x ri D f ) = f(x) i I(x) f i (x). Proof. Let [x, y] (ri D f ) E. By the compatibility in (b) and Proposition 3.5, there exist a colinearly ordered tuple (x 0,..., x n ) X n+1 and functions f i1,..., f in in F such that (36) x 0 = x; x n = y; and ( k {1,..., n} ) [x k 1, x k ] D fik ri D f. 9

10 Define (37) ( k {1,..., n} ) g k := f ik + ι [xk 1,x k ] = f + ι [xk 1,x k ]. Using (34), we see that, for every k {1,..., n 1}, (38) g k (x k ) g k+1 (x k ) f ik (x k ) f ik+1 (x k ) i I(x k ) f i (x k ) =. By applying Theorem 4.4 to the system {g k } k {1,...,n}, we see that g = min k {1,...,n} g k = f + ι [x,y] is convex and hence so is f [x,y]. In view of Lemma 2.6, we obtain the convexity of f. Finally, for all x ri D f, we have f(x) = by Fact 2.3. Therefore, (35) follows from Lemma 4.2. Remark 5.2 We note an interesting feature in Theorem 5.1. In assumption (c), we require the non-emptiness of the subdifferential intersection (34) at all relative interior points except for finitely many points. Since our conclusion says that f is convex, the subdifferential intersection is nonempty at every point in ri D f (see Fact 2.3). That means, in order to check the convexity of f, we are allowed to ignore verifying (34) at finitely many points in ri D f. This turn out to be very convenient since checking (34) at certain points may not be obvious (see, for example, Example 6.1). Remark 5.3 (compatibility on the system of domains is essential) Theorem 5.1 fails if the domain compatibility assumption (b) is omitted: Indeed, Suppose that X = R 2, that I = {1, 2}, and that (39) f 1 = ι [0,1] [0,1] and f 2 = ι [ 1,0[ [0,1] + 1. Then F is a compatible system of functions. Even though D f = [ 1, 1] [0, 1] is convex, {D fi ri D f } i I is not a compatible system of sets. So, Theorem 5.1(b) is violated. Clearly, we can check that f is not convex. Theorem 5.4 (main result II) Assume (23), that F is a compatible system of functions (recall Definition 4.1), that each f i is differentiable on int D fi =, and that the following hold: (a) D f = i I D fi is convex and at least 2-dimensional. (b) {D fi } i I is a compatible system of sets (recall Definition 3.1). (c) There exists a finite subset E of X such that (40) } x (int D f ) E {i, j} I(x) lim z x f i (z) = lim z x z int D fi z int D fj f j (z) exists. Then f is convex; moreover, it is continuously differentiable on(int D f ) E. 10

11 Proof. We will prove the convexity of f by using Theorem 5.1. Note that it suffices to verify assumption (c) of Theorem 5.1. To this end, let x int(d f ) E such that card I(x) 2 and denote by u x the limit in (40). Fact 2.10 and Lemma 4.2 imply (41) u x i I(x) f i (x). So assumption (c) in Theorem 5.1 holds. Thus, we conclude that f is convex. Turning towards the differentiability statement, let Ω be the set of points at which f is differentiable. Then i I int D fi Ω int D f. Now let x (int D f ) E. We consider two cases. Case 1: card I(x) = 1. Then Proposition 3.8 implies that x int D fi for some i I, which implies that x Ω. Case 2: card I(x) 2. Since x int D f, we obtain N D f (x) = {0}. Hence, by Fact 2.10, we have (42) f(x) = cl conv { lim f(z n ) Ω zn x } n N Let (z n ) n N be a sequence in Ω such that z n x, and let (ε n ) n N be in R ++ such that ε n 0 +. By Fact 2.9, f Ω is the continuous and because i I int D fi is dense in D f, there exists a sequence(y n ) n N in i I int D fi such that (43) ( n N) y n z n ε n and f(y n ) f(z n ) ε n. Combining with (40), we deduce that (44) y n x and lim n N f(z n ) = lim n N f(y n ) = u x. So, (42) becomes f(x) = {u x}. Thus, f is differentiable at x by Fact 2.8. Corollary 5.5 Assume (23), that F is a compatible system of functions (recall Definition 4.1), and that the following hold: (a) D f = i I D fi is convex and at least 2-dimensional. (b) {D fi } i I is a compatible system of sets (recall Definition 3.1). (c) f is continuously differentiable on int D f. Then f is convex. Proof. This follows from Theorem 5.4 with E =. 11

12 Remark 5.6 (convex interpolation) When X = R 2, then Corollary 5.5 is known and of importance in the convex interpolation of data; see, e.g., [5, Theorem 1], [6, Proposition 2.2], [12, Theorem 3.1], [2, Proposition 5.1] and the related [13, Theorem 3.6] for further details. Corollary 5.7 Assume (23), that F is a compatible system of functions (recall Definition 4.1), that each D fi is closed, and that ( i I)( x D fi ) f i (x) = 1 2 x, A ix + b i, x +γ i, where A i : X X is linear with A i = Ai 0, b i X, and γ i R. Furthermore, assume that D f = i I D fi is convex and at least 2-dimensional, and that {i, j} I (45) i = j x D fi D A ix+b i = A j x+b j. fj Then f is convex; moreover, it is continuously differentiable on int D f. Proof. This follows from Example 3.2 and Theorem 5.4 with E = since f i (x) = A i x+b i when x int D fi. Remark 5.8 (piecewise linear-quadratic function) Consider Corollary 5.7 with the additional assumption that each D fi is a polyhedral set. Then Corollary 5.7 provides a sufficient condition for checking the convexity of f which in this case is a piecewise linearquadratic function. These functions play a role in computer-aided convex analysis (see [7] and also [11, Section 10.E] for further information on functions of this type). Moreover, we thus partially answer an open question from [8, Section ]. Our results also enhance our understanding of how nonconvexity occurs and form another step towards building a nonconvex toolbox that extends current bivariate computational convex analysis algorithms [3, 4]. 6 Checking convexity We start with an application of Theorem 5.4. Example 6.1 The function x 2 + y max{0, xy} (46) f : R 2, if (x, y) = (0, 0); R : (x, y) x + y 0, if (x, y) = (0, 0) is convex, and differentiable on R 2 {(0, 0)}. Proof. First, set I := {1,..., 4} and { x+y, if (x, y) A f 1 (x, y) := 1 := R 2 + (47a), +, otherwise; 12

13 (47b) (47c) (47d) { x 2 +y 2 f 2 (x, y) := x+y, if (x, y) A 2 := R R +, +, otherwise; { x y, if (x, y) A3 := R 2 f 3 (x, y) :=, +, otherwise; f 4 (x, y) := { x 2 +y 2 x y, if (x, y) A 4 := R + R, +, otherwise. Then { f i } i I is a compatible system of functions (recall Definition 4.1) with f being the corresponding piecewise-defined function. Moreover, {D fi } i I = {A i } i I is a compatible system of sets. Each f i is differentiable on int A i with the gradient given by (48a) (48b) (48c) (48d) f 1 (x, y) = (1, 1) for (x, y) int A 1 ; ( ) x f 2 (x, y) = 2 +2xy+y 2, x2 2xy+y 2 for (x, y) int A (x y) 2 (x y) 2 2 ; f 3 (x, y) = ( 1, 1) for (x, y) int A 3 ; ( ) x f 4 (x, y) = 2 2xy y 2, x2 +2xy y 2 for (x, y) int A (x y) 2 (x y) 2 4. One readily checks that the Hessian of each f i is positive semi-definite on int A i ; hence, by the continuity of f i D fi, we have that f i is convex. Moreover, (49a) (49b) (49c) (49d) ( a > 0) ( a < 0) ( b > 0) ( b < 0) lim f 1 (x, y) = lim f 4 (x, y) = (1, 1); (x,y) (a,0) (x,y) (a,0) (x,y) int A 1 (x,y) int A 4 lim f 2 (x, y) = lim f 3 (x, y) = ( 1, 1); (x,y) (a,0) (x,y) (a,0) (x,y) int A 2 (x,y) int A 3 lim f 1 (x, y) = lim f 2 (x, y) = (1, 1); (x,y) (0,b) (x,y) (0,b) (x,y) int A 1 (x,y) int A 2 lim f 3 (x, y) = lim f 4 (x, y) = ( 1, 1). (x,y) (0,b) (x,y) (0,b) (x,y) int A 3 (x,y) int A 4 Now set E := {(0, 0)}. From the above computations, we observe that all assumptions of Theorem 5.4 are satisfied. Thus, we conclude that f is a convex function that is also continuously differentiable away from the origin. In fact, the function defined by (46) is actually a norm since it is clearly positively homogeneous. An analogous use of Theorem 5.4 allows for a systematic proof of the convexity of the function considered next. Example 6.2 The function (50) f : R 2 R : (x, y) { x6 + y 4, if xy 0; 13 x 3 + y 2, otherwise,

14 is a convex and continuously differentiable. We conclude this section with an application of Theorem 5.1. Example 6.3 The function (51) f : R 2 R : (x, y) f(x, y) := is convex. { x4 + y 2, if xy 0; x 2 + y, otherwise, Proof. First, set I := {1,..., 4} and { x f 1 (x) := 1 4+ x2 2, if x A 1 := R 2 (52a) +, +, otherwise; { x 2 f 2 (x) := 1 + x 2, if x A 2 := R R +, (52b) +, otherwise; { x f 3 (x) := 1 4+ x2 2, if x A 3 := R 2 (52c), +, otherwise; { x 2 f 4 (x) := 1 x 2, if x A 4 := R + R, (52d) +, otherwise. Then { f i } i I is a compatible system of functions (recall Definition 4.1) with f being the corresponding piecewise function. Moreover, {D fi } i I = {A i } i I is a compatible system of sets. Each f i is differentiable on int A i with the gradient given by ( 2x 3 ) f 1 (x) = 1 x (53a), 2 for x int A x 4 1 +x2 2 x 4 1 +x2 2 1 ; (53b) (53c) (53d) f 2 (x) = (2x 1, 1) for x int A 2 ; ( 2x 3 ) f 3 (x) = 1, x 2 for x int A x 4 1 +x2 2 x 4 1 +x2 2 3 ; f 4 (x) = (2x 1, 1) for x int A 4. Next, since the Hessian of f i is positive semidefinite on int A i, we deduce that each f i is convex. Now set E := {(0, 0)}. We will verify (34). Note that simple computations show the following: For x = (0, x 2 ) (A 1 A 2 ) E, (54) lim z x f 1 (z) = lim z x f 2 (z) = (0, 1) f 1 (x) f 2 (x). z int A 1 z int A 2 14

15 For x = (0, x 2 ) (A 3 A 4 ) E, (55) lim z x f 3 (z) = lim z x f 4 (z) = (0, 1) f 3 (x) f 4 (x). z int A 3 z int A 4 For x = (x 1, 0) (A 2 A 3 ) E, (56a) (56b) lim z x f 2 (z) = (2x 1, 1) and N A2 (x) = {0} R ; z int A 2 lim z x f 3 (z) = (2x 1, 0) and N A3 (x) = {0} R +. z int A 3 Then, using Fact 2.10, we conclude that f 2 (x) f 3 (x) =. For x = (x 1, 0) (A 1 A 4 ) E, (57a) (57b) lim z x f 1 (z) = (2x 1, 0) and N A1 (x) = {0} R ; z int A 1 lim z x f 4 (z) = (2x 1, 1) and N A4 (x) = {0} R +. z int A 4 Then, using Fact 2.10, we conclude that f 1 (x) f 4 (x) =. So, we have verified that assumption (c) in Theorem 5.1 holds. Therefore, f is convex by Theorem Detecting the lack of convexity The nonempty subdifferential intersection condition (34) is indeed crucial for the check of convexity: we will see in the following result that the violation of (34) leads to nonconvexity. Theorem 7.1 (detecting lack of convexity) Assume (23), that F is a compatible system of functions (recall Definition 4.1), and that (58) ( x ri D f ) Then f is not convex. i I(x) f i (x) =. Proof. Using Lemma 4.2, we have f(x) i I(x) f i (x) =. Therefore, by Fact 2.3, f is not convex. Using Theorem 7.1, we will now illustrate that the finiteness assumption on E is important for our main results (Theorem 5.1 and Theorem 5.4). 15

16 Example 7.2 Suppose that X = R 2, set { max{ x, y}, if (x, y) A f 1 (x, y) := 1 := R + R; (59a) +, otherwise, { max{x, y}, if (x, y) A 2 := R R; (59b) f 2 (x, y) := +, otherwise, F := { f 1, f 2 }, f := min{ f 1, f 2 }, i.e., (60) f : R 2 R : (x, y) max { x, y }, and E := {0} R. Then one checks the following: (i) F is a compatible system of functions. (ii) {A 1, A 2 } is a compatible system of sets. (iii) For every(x, y) R 2 E with I F (x, y) = {1, 2}, we must have (x, y) {0} R ++, i.e., x = 0 and y > 0. Then f(x, y) = y locally around (0, y) and thus (61) f 1 (0, y) f 2 (0, y) f(0, y) = {(0, 1)}. So, all assumptions in Theorems 5.1 and 5.4 are satisfied except that E is infinite. However, for (0, y) E {(0, 0)}, we have y < 0; thus, f 1 (x, y) = x+ι A1 (x, y) locally around (0, y). It follows that (62) f 1 (0, y) = ( 1, 0)+N A1 (0, y) = ], 1] {0} and similarly that (63) f 2 (0, y) = (1, 0)+ N A2 (x, y) = [1,+ [ {0}. Hence f 1 (0, y) f 2 (0, y) = and so f is not convex by applying Theorem 7.1 or by direct inspection. In the previous example, the set E was infinite, but unbounded. In the next (slightly more involved) example, we provide a case where E is bounded. Example 7.3 Suppose that X = R 2, set I := {1,..., 6}, (64a) (64b) (64c) (64d) A 1 := { (x, y) R 2 x + y 1, x 0, y 0 }, A 2 := { (x, y) R 2 x + y 1, x 0, y 0 }, A 3 := { (x, y) R 2 x + y 1, x 0, y 0 }, A 4 := { (x, y) R 2 x + y 1, x 0, y 0 }, 16

17 (64e) (64f) A 5 := { (x, y) R 2 x + y 1, x 0 }, A 6 := { (x, y) R 2 x + y 1, x 0 }, (65) f : R 2 R : (x, y) max { 1 x, y }, (66) F := { f i } i I, where ( i I) f i := f + ι Ai, and E := ( {0} [ 1, 1] ) {(±1, 0)}. Then one checks the following: (i) Each f i is convex and continuous on D fi because (67a) (67b) (67c) (67d) f i (x, y) = y+ι Ai (x, y) for i {1, 2}; f i (x, y) = y+ι Ai (x, y) for i {3, 4}; f 5 (x, y) = x+1+ι A5 (x, y); and f 6 (x, y) = x+ 1+ι A6 (x, y). Consequently, F is a compatible system of functions. (ii) {A i } i I is a compatible system of sets. (iii) f is the piecewise-defined function associated with F. (iv) Take (x, y) R 2 E with card I F (x, y) 2. Then (68) I F (x, y) { {1, 2},{1, 4},{1, 5},{2, 3},{2, 6},{3, 4},{3, 6},{4, 5} }. Suppose, for instance, that I F (x, y) = {1, 5}. Then x > 0, y > 0, and x+y = 1. We have (69) f 1 (x, y) = (0, 1)+ N A1 (x, y) = (0, 1)+R + ( 1, 1) and (70) f 5 (x, y) = (0, 1)+ N A5 (x, y) = ( 1, 0)+R + (1, 1). Then (71) ( 1 2, 1 2 ) f 1(x, y) f 5 (x, y); similarly, one obtains nonemptiness for the other cases. We observe that all assumptions in Theorem 5.1 and Theorem 5.4 are satisfied except that E is infinite and bounded. However, for every (0, y) {0} ] 1, 1[ E, we have f(x, y) = min{ f 5 (x, y), f 6 (x, y)} = x +1 locally around(0, y). Clearly, f is not convex. 17

18 Acknowledgments HHB was partially supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada and by the Canada Research Chair Program. YL was partially supported by the Natural Sciences and Engineering Research Council of Canada through a Discovery grant. HMP was partially supported by an NSERC accelerator grant of HHB. References [1] H.H. Bauschke and P.L. Combettes, Convex Analysis and Monotone Operator Theory in Hilbert Spaces, Springer, [2] W. Dahmen, Convexity and Bernstein-Bézier polynomials, in Curves and surfaces (Chamonix-Mont-Blanc, 1990), pages , Academic Press, [3] B. Gardiner, K. Jakee, and Y. Lucet, Computing the partial conjugate of convex piecewise linear-quadratic bivariate functions, Computational Optimization and Applications 58 (2014), [4] B. Gardiner and Y. Lucet, Computing the conjugate of convex piecewise linearquadratic bivariate functions, Mathematical Programming (Series B) 139 (2013), [5] T.A. Grandine, On convexity of piecewise polynomial functions on triangulations, Computer Aided Geometric Design 6 (1989), [6] A. Li, Convexity preserving interpolation, Computer Aided Geometric Design 16 (1999), [7] Y. Lucet, What shape is your conjugate? A survey of computational convex analysis and its applications, SIAM Review 52 (2010), [8] Y. Lucet, Techniques and open questions in computational convex analysis, in Computational and Analytical Mathematics, pages , Springer, [9] B.S. Mordukhovich and N.M. Nam, An easy path to convex analysis and applications, Morgan & Claypool, [10] R.T. Rockafellar, Convex Analysis, Princeton University Press, [11] R.T. Rockafellar and R.J-B. Wets, Variational Analysis, Springer, [12] L.L. Schumaker and H. Speleers, Convexity preserving splines over triangulations, Computer Aided Geometric Design 28 (2011), [13] L.L. Schumaker and H. Speleers, Convexity preserving C 0 splines, Advances in Computational Mathematics 40 (2014), [14] C. Zălinescu, Convex Analysis in General Vector Spaces, World Scientific,